acceptodds
Under review as a conference paper at ICLR 2027

HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses

Abstract

Scientific agents increasingly contribute to hypothesis discovery by synthesizing evidence, assessing proposals, and developing new explanations. Recent systems bring scientific agents and evolutionary search together to develop hypotheses through cycles of critique, comparison, and revision. However, how different forms of agent collaboration affect hypothesis quality remains an open question. Answering this question requires separating the effects of agents' scientific capabilities from those of their collaboration. A suitable framework must therefore preserve the agents' scientific roles and support different rules for combining, revising, and retaining hypotheses. Building on this perspective, we introduce HypoEvolve, which makes collaboration explicit through successive updates to a hypothesis population. Specifically, we propose to use a generational genetic algorithm to coordinate specialized large language model (LLM) agents that integrate mechanistic arguments, reconsider assumptions, and assess evidence and testability. Each generation specifies how scientific judgments and new proposals reshape the population, which makes the effects of collaboration on hypothesis quality directly testable. Moreover, we design our evaluation around scientifically meaningful hypotheses that explain how a proposed intervention could work. Drug repurposing connects these explanations to target-level biological claims that can be assessed against external evidence. Specifically, we adapt DepMap and Open Targets into complementary external measures grounded in experimental, genetic, and clinical evidence. The evaluation spans 34 cancer types, with HypoEvolve achieving the highest scores against six baselines on both measures. DepMap selectivity reaches 0.171, compared with 0.115 for the strongest baseline. Gains over single-pass generation also generalize to held-out cancer types. HypoEvolve advances a vision of autonomous science in which AI research teams achieve a capacity for discovery beyond that of individual models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific discovery, hypothesis generation, LLM agents, multi-agent systems, evolutionary search, genetic algorithms, drug repurposing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QHLCOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HypoEvolve: a generational genetic algorithm coordinating specialized LLM agents (generation, pairwise comparison, crossover/mutation, etc.) to cast hypothesis refinement as fixed-size population search, with fitness from pairwise judgments driving selection and replacement. The core claim is that, with scientific roles and prompts fixed, collaboration/search rules can themselves be experimental variables. Validation focuses on a closed-set drug-repurposing setting with external biomedical association resources, plus comparisons to Tree-of-Thoughts and single-pass generation, and selection-related ablations.

Critical strengths
  1. Trying to decouple collaboration rules from scientific roles, and contrasting fitness-guided selection against weaker controls, is more methodologically intentional than stacking agent workflows; this is a useful framing for whether organization contributes independently.
  2. The external-evaluation protocol is comparatively careful: answers are submitted before scoring, external evidence does not enter fitness, and held-out settings reduce pure LLM self-scoring circularity.
Critical weaknesses
  1. The evaluation mainly measures biological associations implied by hypotheses, not the full scientific quality of those hypotheses. The paper claims to assess scientific hypothesis discovery with mechanistic rationales, yet the main results largely reduce to DepMap selectivity and Open Targets association on a closed drug set—closer to scoring whether chosen drug–cancer target links match external resources than to directly testing whether rationales are correct, novel, or mechanistically falsifiable.
  2. There is a risk that pairwise fitness judging is aligned with the external metrics, further weakening the discovery narrative. The scoring appears to emphasize cancer-specific dependency and target-evidence cues close to DepMap selectivity, raising the risk of using a nearby criterion as the search referee and then claiming discovery on similar external metrics; the appendix constant-drug control can also trade wins and losses with the main method on held-out settings, weakening the “evolution produced better hypotheses” interpretation.
Questions for authors
  1. After matching call counts and candidate-pool size, do advantages over main search baselines and simpler controls remain stable?
  2. If pairwise scoring is decoupled from external biomedical metrics (e.g., emphasizing mechanistic coherence and testability), do conclusions change?
  3. Does current evidence better support “improved drug–setting matching” or “improved scientific hypothesis discovery,” and how should the writing be scoped?
2 comments
AuthorsOpenReview ↗

Response to Reviewer QHLC (1/2)

Thank you for the review. Your three questions go to the core of the paper, so we ran the experiments they call for. The new runs use a second model, GPT-5.6 Luna, across all 34 cancer types. Each method has one run per cancer type, and single-pass generation has six draws. Scores are means over cancer types (31 for DepMap selectivity, 34 for Open Targets), and p-values are uncorrected unless stated. We answer the questions first.

Q3. Matching or discovery. The evidence supports drug-cancer matching in a narrow form. It does not support hypothesis discovery. We withdraw that claim and the collaboration framing. We will revise the title to match.

The result that holds on both models is higher Open Targets association than single-pass generation. On all 34 cancer types, the margin is +0.273 on gpt-5.4-mini (the model of the paper) and +0.253 on GPT-5.6 Luna, each with p<0.001.

To test matching directly, we use the cancer-specificity residual of Appendix C.4. It subtracts the mean score of the same drug over the other cancer types, so a drug that scores well everywhere gains nothing. On this residual, HypoEvolve leads single-pass generation by +0.089 on DepMap (p=0.014) and +0.092 on Open Targets (p=0.0055) on gpt-5.4-mini. On Luna the lead is +0.085 (p=0.084) and +0.086 (p=0.036).

Q1. Matched calls and candidate pool. The advantages do not remain stable.

MethodLLM callsCandidatesDepMapOpen TargetsMargin, DepMapMargin, Open Targets
Tree of Thoughts, matched candidates81240.0480.304+0.072 (p=0.043)+0.097 (p=0.10)
Tree of Thoughts, matched calls193720.0410.231+0.079 (p=0.003)+0.169 (p=0.010)
Static reranking of 24 candidates308240.0690.269+0.051 (p=0.16)+0.132 (p=0.014)
HypoEvolve206240.1200.400

A margin is the score of HypoEvolve minus the score of the method. We fixed 16 tests before the runs and applied Holm correction. Of these six margins, only the DepMap margin at 193 calls stays below 0.05 after correction (0.044). This result is fragile. Tree of Thoughts often ends with several leaves tied on the top score and returns the first. When we average the tied leaves, the corrected value is 0.18. Tokens are not matched in any comparison.

AuthorsOpenReview ↗

Response to Reviewer QHLC (2/2)

Q2. A judge decoupled from the metrics. We ran the judge you describe. It ranks on mechanistic coherence, evidence and testability, and it does not mention cancer specificity. The conclusions change in part.

With this judge, HypoEvolve scores 0.103 on DepMap and 0.316 on Open Targets. With our judge (the pairwise scorer in the paper) it scores 0.120 and 0.400, and single-pass generation scores 0.026 and 0.147. The generic-judge arm still leads single-pass generation on Open Targets (+0.169, corrected p=0.011). The difference between the two judges is not resolved (+0.017 on DepMap, p=0.69; +0.084 on Open Targets, p=0.064).

However, three of the four operator prompts also ask for cancer specificity, so this arm keeps the cancer-specificity criterion in the search. We therefore ran a further arm with the generic judge and with operator prompts that do not ask for cancer specificity. It scores 0.037 and 0.200, and it is not shown to beat single-pass generation (+0.011, p=0.87; +0.053, p=0.44). The full system leads it on Open Targets by +0.201, with a corrected p-value of 0.005 over the ten tests of this arm.

This arm also changes which drugs the operators propose, so it does not isolate the wording. In short, the Open Targets lead survives a generic judge. We can no longer show it once the criterion is also removed from the operator prompts. Our reply to Reviewer CYkt gives the details.

W1. What the evaluation measures. We agree. Both metrics score the annotated targets of the proposed drug on a closed list of 61 drugs. We do not assess the rationale itself. We will state this as a limitation.

W2. Alignment of the judge with the metrics. The risk you describe is real, and it applies beyond the judge. The judge and three of the four operator prompts ask for cancer specificity, and the single-pass prompt does not. When we put the criterion of the judge into the single-pass prompt, its scores rise from 0.026 to 0.062 on DepMap and from 0.147 to 0.265 on Open Targets. That recovers about half of the Open Targets margin. HypoEvolve leads this prompt by +0.058 on DepMap (p=0.66) and +0.135 on Open Targets (p=0.021, or 0.17 after correction). On Luna, we therefore cannot show that the search adds to the criterion.

On the constant-drug control, we agree that it weakens a reading based on raw scores. The residual described under Q3 is our answer, because any constant drug scores zero on it. The final answers of HypoEvolve score +0.103 on DepMap and +0.100 on Open Targets on gpt-5.4-mini (p=0.018 and p=0.0086 against zero). On Luna they score +0.080 and +0.101 (p=0.16 and p=0.014).

Reviewer_CYktOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents HypoEvolve, a framework that uses a genetic algorithm to coordinate LLM-based hypothesis generation, evaluation, crossover, and mutation. It evaluates the method on cancer drug repurposing across 34 cancer types, selecting from 61 predefined drugs and using DepMap selectivity and Open Targets association as evaluation metrics. The authors report improvements over six baselines and conduct experiments on search configurations and component contributions.

Critical strengths
  1. The paper considers how search and coordination mechanisms affect LLM-based scientific hypothesis generation, which is a relevant research question.
  2. The paper specifies its search procedure and includes comparisons with existing methods, external biological metrics, and component ablations.
Critical weaknesses
  1. Mismatch between the research question and the actual contribution. The paper frames its contribution around multi-agent collaboration, but the method is primarily an evolutionary search algorithm with LLM-based operators. The manipulated variables are GA search rules rather than collaboration mechanisms. Moreover, the main comparisons vary multiple factors simultaneously, so the stated goal of disentangling agent capabilities from collaboration is not adequately addressed.
  2. The source of the performance gains is unclear. The task-specific judge explicitly favors cancer-specific dependencies and target evidence, which closely align with the external metrics, but its contribution is not separated from that of evolutionary search. The core crossover and mutation operators show no reliable benefit in ablations: none of the operator knockouts is statistically significant after correction, and removing both operators even improves DepMap selectivity.
  3. No validated benchmark and narrow experimental scope. The paper does not use an established benchmark for scientific hypothesis discovery, instead constructing its own evaluation without sufficiently validating whether it reflects hypothesis quality. The metrics score only the selected drug rather than the quality of the hypothesis, and a constant-drug baseline even outperforms HypoEvolve on held-out Open Targets. The experiments are also confined to a single task and a single LLM backbone.
  4. Baseline fairness and experimental validation are insufficient. The Co-Scientist baseline excludes its original evolution operators, while HypoEvolve uses similar operations. It is also unclear whether baselines receive the same task-specific judging criteria.
  5. Limited novelty and generality. The method largely builds on existing evolutionary search approaches, without clearly establishing a substantial methodological contribution. The experiments are limited to a single model and task, with generalization tested only across cancer types.
Questions for authors

see weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer CYkt (1/2)

Thank you for the review. We agree with most of it. To test your points, we ran new experiments on a second model, GPT-5.6 Luna, across all 34 cancer types. Each method has one run per cancer type, and single-pass generation has six draws. We withdraw the claims of hypothesis discovery and of multi-agent collaboration. The claim that holds on the second model is narrower, namely higher Open Targets association than single-pass generation. We will revise the title to match.

W1. Framing. We agree. The paper varies search rules, and we will drop the collaboration framing from the abstract and the introduction. We also agree that the main comparisons change several factors at once. One of them is where each method asks for cancer specificity, which we show under W2.

W2. Source of the gains. Your concern about the judge (the pairwise scorer in the paper) is justified, and it applies beyond the judge. The table shows where the prompts of each method ask for cancer specificity.

MethodWhere the prompts ask for cancer specificity
Single-pass generation, self-consistencyNowhere
Multi-agent debateOne sentence in the revise prompt
Static reranking, co-scientist scaffoldThe pairwise judge
Tree of ThoughtsThe evaluator prompt
HypoEvolveThe pairwise judge and three of the four operator prompts

We then ran three new arms on Luna to separate the cancer-specificity criterion from the search. Scores are means over cancer types (31 for DepMap selectivity, 34 for Open Targets), and p-values are uncorrected unless stated.

MethodDepMapOpen Targets
Single-pass generation0.0260.147
Single-pass generation with the criterion0.0620.265
HypoEvolve without the criterion in the judge or the operator prompts0.0370.200
HypoEvolve with a generic judge0.1030.316
HypoEvolve0.1200.400

First, the criterion in one prompt recovers about half of the Open Targets margin. We put the criterion of the judge into the single-pass prompt, in place of its instruction to describe a broadly applicable mechanism. Open Targets rises from 0.147 to 0.265 (p=0.004). HypoEvolve leads this prompt by +0.058 on DepMap (p=0.66) and +0.135 on Open Targets (p=0.021, or 0.17 after Holm correction over the 16 tests we fixed before the runs). We therefore cannot show that the search adds to the criterion.

Second, a generic judge does not remove the lead. This judge ranks on mechanistic coherence, evidence and testability. With it, HypoEvolve still leads single-pass generation on Open Targets (+0.169, corrected p=0.011). However, its operator prompts still ask for cancer specificity.

Third, we can no longer show a lead when the criterion is removed from the operator prompts as well. This arm is not shown to beat single-pass generation (+0.011 on DepMap, p=0.87; +0.053 on Open Targets, p=0.44). The full system leads it on Open Targets by +0.201. The corrected p-value is 0.005 over the ten tests of this arm, which we fixed before it ran and after we had seen the first results. On the cancer-specificity residual of Appendix C.4 (described under W3), that lead is +0.069 (p=0.093).

We conclude that the Open Targets gain of HypoEvolve depends on the cancer-specificity criterion in its prompts. One limit applies to the third arm. Without the criterion, the inspiration operator proposes non-oncology drugs in 73% of its offspring, against 10% with it. The arm therefore changes the proposals as well as the wording.

AuthorsOpenReview ↗

Response to Reviewer CYkt (2/2)

W2. Operator knock-outs. We agree that the knock-outs on eight cancer types show no reliable benefit. The DepMap increase after removing both operators (+0.034) comes from one cancer type, head and neck. Without it, the change is -0.020. Open Targets falls by 0.121.

Every run also contains a larger version of this knock-out. The answer a run would return from its initial population is its result with both operators off. On Luna, the final answer beats it by +0.100 on DepMap (corrected p=0.051) and by +0.161 on Open Targets (corrected p=0.012). In the arm without the criterion, it does not (-0.015 and -0.077). This comparison changes operators, selection rounds and the number of candidates together, so it does not separate the operators.

W3. Benchmark, metrics and scope. We agree on each point. We use no established benchmark. Both metrics score the annotated targets of the proposed drug, and nothing assesses the rationale itself.

On the constant-drug baseline, we agree that raw association scores reward a well-annotated drug. The cancer-specificity residual of Appendix C.4 removes this effect. It subtracts the mean score of the same drug over the other cancer types, so any constant drug scores zero. On this residual, the final answers of HypoEvolve score +0.103 on DepMap and +0.100 on Open Targets on gpt-5.4-mini, the model of the paper (p=0.018 and p=0.0086 against zero). On Luna they score +0.080 and +0.101 (p=0.16 and p=0.014).

On the backbone, we now have results on three models from one vendor. They are gpt-5.4-mini, gpt-5.4 (run before submission and not reported in the paper) and GPT-5.6 Luna. On all scored cancer types, the Open Targets margin over single-pass generation is +0.273, +0.206 and +0.253, each with p<0.001. The DepMap margin is +0.12 (p<0.001), +0.045 (p=0.071) and +0.094 (p=0.043, or 0.30 after correction). The DepMap result is therefore firm on gpt-5.4-mini only. We have no second task, and we limit our claims to drug repurposing.

W4. Baselines. The co-scientist scaffold keeps only generation, ranking and meta-review. It is not a full Co-Scientist, and we have no comparison with the full system. We will state this where the baseline is introduced.

The baselines do not all receive the criterion of the judge, as the first table in comment 1/2 shows. Single-pass generation with the criterion is the direct test, and it closes about half of the Open Targets gap.

W5. Novelty and generality. We agree. The search uses standard evolutionary mechanisms, and we claim no methodological advance. The evidence covers one task.

Reviewer_4yB5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HypoEvolve, a framework that coordinates LLM agents through a generational genetic algorithm to develop scientific hypotheses. Agents perform literature-grounded generation, pairwise comparison, and semantic crossover and mutation. Bradley–Terry scores guide tournament selection and joint parent–offspring replacement in a fixed-size population. The evaluation concerns drug repurposing across 34 cancer types using a constrained vocabulary of 61 drugs. Hypotheses are selected before external assessment with DepMap selectivity and Open Targets association. HypoEvolve achieves the highest mean scores among the evaluated methods on the shared evaluation panels, and improvements over single-pass generation persist on held-out cancer types. A controlled parent-selection experiment also shows improvements in population-average and minimum external scores.

Critical strengths
  1. The evaluation separates search-time judgments from external biological assessment. DepMap and Open Targets scores are applied only after each method commits to an answer. This provides more meaningful evidence than improvements in the system’s own LLM-assigned fitness. The selectivity adjustment also addresses an important failure mode of raw dependency scores: rewarding targets that are broadly essential across cancers.

  2. The framework enables a useful controlled study of search organization. The parent-selection experiment changes fitness-guided selection to uniform selection while holding the scientific operators and hypothesis count fixed. Its improvements in population mean and minimum scores provide concrete evidence that a coordination rule can affect candidate quality. The paper also reports informative negative results and limitations, including inconclusive operator ablations and mixed effects of additional search.

Critical weaknesses
  1. The methodological novelty is incremental, and the distinctive components are not sufficiently supported by the ablations. The paper’s own related-work discussion identifies prior systems combining scientific agents with hypothesis evolution, population search, and semantic crossover. HypoEvolve’s main distinction is the particular combination of tournament selection, fixed-size generations, and joint parent–offspring truncation. This is a clear implementation of established evolutionary mechanisms, but the evidence does not yet establish why this combination constitutes a substantial advance. In Section 4.4, removing crossover or mutation produces mixed effects across eight cancer types, and no operator-removal comparison survives multiple-testing correction. These results do not establish that the operators are ineffective, but they leave their central contribution unresolved. The strongest controlled finding concerns population mean and minimum quality; population maxima show no statistically detectable improvement.

  2. The evaluation supports improved target–cancer matching more directly than improved scientific discovery, and the comparative advantage remains qualified. The external metrics score annotated drug targets and cancer associations rather than directly testing the generated mechanistic explanations or their novelty. Appendix C.4 makes this limitation concrete: a constant-drug control exceeds HypoEvolve on held- out Open Targets association, although DepMap favors HypoEvolve. Thus, higher association scores alone do not establish better cancer-specific hypothesis discovery. Against the strongest baseline, Tree of Thoughts, the DepMap improvement is statistically significant, but the Open Targets difference is unresolved ((p=0.126)). HypoEvolve also uses 206 model calls per run versus 31 for Tree of Thoughts, without matched token budgets. Taken together, these results demonstrate useful performance in a constrained application, but do not yet clearly isolate the benefit of the proposed search design over a comparably resourced alternative.

Questions for authors
  1. Clarification of methodological contribution: Which specific contribution beyond prior scientific-agent evolution and semantic crossover frameworks do the authors consider central? What existing experimental evidence isolates the benefit of fixed-size generational replacement, independently of parent selection and the variation operators?

  2. Clarification of evaluation validity: Is there any assessment of the generated rationale itself—its mechanistic correctness, novelty, or falsifiable predictions—beyond the selected drug’s annotated target associations? This would materially affect whether the results support hypothesis discovery or primarily improved drug–cancer prioritization.

  3. Additional evidence affecting the recommendation: Do the authors have comparisons with Tree of Thoughts or static reranking under matched model-token budgets and comparable candidate counts? Such results would help determine whether the improvement comes from the proposed organization of search or additional computational resources.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4yB5 (1/2)

Thank you for the detailed review. We ran the comparisons you asked for on a second model, GPT-5.6 Luna, across all 34 cancer types. Each method has one run per cancer type, and single-pass generation has six draws. p-values are uncorrected unless stated. We agree with your main points. We withdraw the claim of hypothesis discovery, and we no longer claim an advantage over Tree of Thoughts at a comparable budget. The claim that holds on the second model is narrower, namely higher Open Targets association than single-pass generation. We will revise the title to match.

Q3 and W2. Comparison at a matched budget. We now have these comparisons, at matched calls and at matched candidates. On the second model, they show no reliable advantage of HypoEvolve over Tree of Thoughts or static reranking.

MethodLLM callsCandidatesInput tokensDepMapOpen TargetsMargin, DepMapMargin, Open Targets
Single-pass generation111K0.0260.147+0.094 (p=0.043)+0.253 (p<0.0001)
Tree of Thoughts, setting of the paper3195K0.1040.289+0.016 (p=0.53)+0.112 (p=0.048)
Tree of Thoughts, matched candidates812412K0.0480.304+0.072 (p=0.043)+0.097 (p=0.10)
Tree of Thoughts, matched calls1937225K0.0410.231+0.079 (p=0.003)+0.169 (p=0.010)
Static reranking30824405K0.0690.269+0.051 (p=0.16)+0.132 (p=0.014)
HypoEvolve20624257K0.1200.400

Scores are means over cancer types (31 for DepMap selectivity, 34 for Open Targets). A margin is the score of HypoEvolve minus the score of the method, with an uncorrected p-value from a paired Wilcoxon test. Static reranking ranks 24 candidates once with the same pairwise scorer, so it matches candidates only. Tokens are not matched in any comparison.

We fixed 16 tests before the runs and applied Holm correction. The Open Targets margin over single-pass generation holds (corrected p=0.0007). The DepMap margin over single-pass generation does not (0.30). Of the six margins over Tree of Thoughts at 81 and 193 calls and over static reranking, only the DepMap margin at 193 calls stays below 0.05 after correction (0.044). This result is fragile. Tree of Thoughts often ends with several leaves tied on the top score and returns the first. When we average the tied leaves, the corrected value is 0.18.

As you note, the Open Targets difference from Tree of Thoughts was unresolved in the paper (p=0.126). On the second model, the DepMap margin at the setting of the paper is not reproduced either (+0.016).

AuthorsOpenReview ↗

Response to Reviewer 4yB5 (2/2)

Q1. Central contribution and evidence for fixed-size replacement. No experiment isolates fixed-size replacement. The selection study switches fitness off in parent selection and in replacement at once, so it cannot separate the two. As you note, it also shows no gain in the best member of the population.

We see two contributions. The first is the evaluation protocol. Every method commits to one answer before scoring, and the protocol includes the selectivity adjustment and the constant-drug control. The second is an empirical finding. The final answer of a run scores higher than the answer the same run would return from its initial population, before any crossover or mutation. On gpt-5.4-mini, the model of the paper, the margins are +0.119 on DepMap (p=0.0010) and +0.151 on Open Targets (p=0.0003). On Luna they are +0.100 (p=0.004) and +0.161 (p=0.0008). After correction the Luna values are 0.051 and 0.012, so we state this finding for Open Targets only. In 28 of 34 Luna runs, the final drug does not appear in the initial population.

This comparison changes operators, selection rounds and the number of candidates together, so it does not show which component helps. Our reply to Reviewer CYkt explains how much of it depends on the cancer-specificity criterion in our prompts.

W1. Novelty. We agree. Tournament selection, fixed-size generations and joint parent-offspring replacement are standard evolutionary mechanisms. We claim no methodological advance over the prior systems you cite.

W1. Operator ablations. We agree. Eight cancer types cannot separate crossover from mutation, and we do not have a larger ablation.

W2 and Q2. What the metrics measure. We agree. Both metrics score the annotated targets of the proposed drug. We do not assess the rationale itself, so we cannot speak to its correctness, novelty or predictions. We will state this as a limitation, and we withdraw the claim of hypothesis discovery.

W2. Constant-drug control. We agree that higher association scores alone do not show cancer-specific matching. The cancer-specificity residual of Appendix C.4 addresses this point. It subtracts the mean score of the same drug over the other cancer types, so any constant drug scores zero. On this residual, the final answers of HypoEvolve score +0.103 on DepMap and +0.100 on Open Targets on gpt-5.4-mini (p=0.018 and p=0.0086 against zero). On Luna they score +0.080 and +0.101 (p=0.16 and p=0.014).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.